Everyone talks about taste. In an era where AI can execute almost every step of research—writing code, surveying papers, running or even designing experiments—the central question is no longer how to solve problems, but which problems deserve to be solved. What makes a problem meaningful in the first place?

I just graduated and am taking my first steps in research. I know taste matters, and I want to build my own. But the harder question is: how do I actually develop it?

One way to think about taste is as choosing a seed node in a vast search space. Research is not a well-defined optimization problem. There is no clear objective, no ground truth, no reliable gradient to follow. Instead, we are placed in a large, unstructured space and forced to decide where to begin. Taste, in this sense, is not just the ability to pick good starting points, but to repeatedly make decisions about direction: to pursue, to pivot, or to abandon. It is the ability to steer toward regions where something meaningful might emerge, even when the path itself is uncertain.

This is where current AI systems show their limits. They perform remarkably well once the search space is specified. Given a seed, they can explore efficiently and learn from the resulting trajectories, but they largely learn from trajectories after the search has already been structured. In reality, many important discoveries arise not from optimizing within a known space, but from wandering through an undefined one—through repeated trial and error, false starts, and dead ends. Most of this process is never recorded, structured, or turned into training data.

This suggests that developing taste is not just about choosing problems, but about building intuition for what makes a problem worth choosing in the first place. And that intuition can only come from direct exploration.

There is no ground truth to guide this. Yet we are constantly required to make decisions: which directions to pursue, which to abandon, where to focus. Why does one problem feel meaningful while another does not? These judgments are not the output of explicit optimization. They are formed through experience. Developing taste, in the end, is the process of constructing an internal objective function: one that is never explicitly defined, but gradually emerges through exploration in a vast and ever-expanding space.

If taste is built through direct exploration, it might seem like something that can only be learned through experience, iteration, and failure. While this kind of firsthand exploration is important, I think it can also be developed indirectly through proxies. My favorite way to do this is also the simplest: reaching out and talking to people. Conversations let me step outside my own thinking and extrapolate beyond what I could discover alone. I try to understand how others formed their sense of taste.

What are the most important problems in your field, and why are you not working on them? Or more simply, what feels worth your time right now? I also try to look across domains. What problems are people excited about? What remains unsolved? When I talk to people who are further along, I am especially interested in how their perspective differs from mine, what they see that I cannot yet see.

This started before I graduated, when I was trying to decide whether to pursue a PhD. I wanted to build my own solid reasons for that decision, so I started having one-on-one conversations, nearly a hundred of them. Through these conversations, I encountered out-of-the-box thinking I could not have produced alone, and possibilities I did not know existed. I learned how much it matters to keep the door open, keep asking, and keep listening.

More importantly, I realized that I genuinely enjoy this process. It felt like sending a query and distilling the other person’s mental model through their answer. This is also why I keep thinking about good questions. What makes a question worth asking? Which questions lead somewhere, and which ones collapse quickly? In some sense, this feels similar to the problems I study in information retrieval, searching and exploring over a vast space with limited signals.

Now I think I am at the stage of figuring out what my own questions are, how to work through them, and how to convince others that those questions and answers are worth taking seriously. Ultimately, I want to find my own good questions—the ones I find genuinely interesting—and engage with them directly. The parts of this process that are never written down or learnable by others are exactly what I want to find out for myself.

References

My previous essays: Mental Models (Nov 2025), Optimization in Life (Jun 2024)

Thanks to my friends, whose thoughtful feedback considerably improved this piece.